[NumPy-Tickets] [NumPy] #2209: Inconsistent dot product with BLAS

[NumPy Trac]

#2209: Inconsistent dot product with BLAS
------------------------------------------------+---------------------------
 Reporter:  law                                 |       Owner:  somebody   
     Type:  defect                              |      Status:  new        
 Priority:  normal                              |   Milestone:  Unscheduled
Component:  numpy.core                          |     Version:  1.6.2      
 Keywords:  dot product blas dotblas sse2 sse3  |  
------------------------------------------------+---------------------------
 Hello,

 I found some strang bug related to numpy dot product. When using numpy-
 sse2 or numpy-sse3 I get different results for the same data and the same
 script. The test script is as follows:

 {{{
 #test script
 from numpy import *

 a = [-0.07738518658898276004,  0.96881291391416879399,
 -0.00769691931712315151, 0.08111387479193514494, 0.11885744586034743819,
 0.05882337083773133551]
 a2 = [ \
   [0, 0, 0, 0, 0, 0], \
   [-0.07738518658898276004,  0.96881291391416879399,
 -0.00769691931712315151, 0.08111387479193514494, 0.11885744586034743819,
 0.05882337083773133551] \
   ]

 b = [0.89887845358462115097, -0.08609082382649151099,
 -0.02124706755589483506, 0.22391191935591209239,  0.32810168298084929628,
 0.16237978892078003268]

 set_printoptions(precision=20)

 ab = dot(a, b)
 a2b = dot(a2, b)

 ab_i = inner(array(a), swapaxes(array(b), -1, -2))
 a2b_i = inner(array(a2), swapaxes(array(b), -1, -2))

 r = 0
 for i in range(6):
     r = r + a[i] * b[i]

 print "dot: %.20f" % ab
 print "dot2: %.20f" % a2b[1]
 print "inner: %.20f" % ab_i
 print "inner2: %.20f" % a2b_i[1]
 print "sum_prod: %.20f" % r
 }}}


 And I get:

 {{{
 C:\Documents and Settings\1>python c:\temp\t0.py
 dot: -0.08609082382649152487
 dot2: -0.08609082382649149712
 inner: -0.08609082382649152487
 inner2: -0.08609082382649149712
 sum_prod: -0.08609082382649149712

 C:\Documents and Settings\1>python -i c:\temp\t0.py
 dot: -0.08609082382649152487
 dot2: -0.08609082382649151099
 inner: -0.08609082382649152487
 inner2: -0.08609082382649151099
 sum_prod: -0.08609082382649149712
 >>> quit()

 C:\Documents and Settings\1>python c:\temp\t0.py
 dot: -0.08609082382649152487
 dot2: -0.08609082382649149712
 inner: -0.08609082382649152487
 inner2: -0.08609082382649149712
 sum_prod: -0.08609082382649149712

 C:\Documents and Settings\1>python -i c:\temp\t0.py
 dot: -0.08609082382649152487
 dot2: -0.08609082382649151099
 inner: -0.08609082382649152487
 inner2: -0.08609082382649151099
 sum_prod: -0.08609082382649149712
 >>> quit()
 }}}


 Surely, I don't change the script or anything else between runs. It seems
 that the result depends on something, but I can't figure out what exactly.
 In this particular example that difference caused by -i switch, other time
 result changed when I add -x or -u, or just run script two time in a row.
 I come to conclusion that this strange behavior related to dotblas, cause
 I tried to replace numpy-sse3 with numpy-1.6.2-nosse (numpy-1.7.0b1-nosse)
 the results are constant and match to simple sum of products:
 C:\Documents and Settings\1>python c:\temp\t0.py
 dot: -0.08609082382649149712
 dot2: -0.08609082382649149712
 inner: -0.08609082382649149712
 inner2: -0.08609082382649149712
 sum_prod: -0.08609082382649149712

 At the same time I see that "random" fluctuations with numpy-1.6.2-sse2,
 numpy-1.6.2-sse3, numpy-1.7.0b1-sse2, numpy-1.7.0b1-sse3 and on Windows
 XP, Windows 2008, Windows 7 (tested on real and virtual machines (AMD
 CPU))

 I can perform other tests if it can be helpful.

 Best regards,
 Alexander

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/2209>
NumPy <http://projects.scipy.org/numpy>
My example project